Publications

2026

1. ICLR 2026

   

   Understanding and Improving Length Generalization in Hierarchical Sparse Attention Models

   Jiaqi Leng^*, Xiang Hu^*, Junxiong Wang, Jianguo Li, Wei Wu, and Yucheng Lu

   In Proceedings of the 14th International Conference on Learning Representations, 2026

   Abs arXiv Code Poster

   We analyze chunk-based sparse attention for long-context generalization and identify its essential ingredients.

   • Key principles:
     • non-linear Chunk Encoder with a dedicated CLS token for retrieval
     • Bypassing Residual Path to integrate global information
     • enforced selection sparsity in pre-training to close the train–test gap
   • Evidence: theoretical motivation and unified ablations
   • Result: state-of-the-art training-free length extrapolation, generalizing 4K-trained models to 32M tokens on RULER and BABILong.
2. Preprint

   

   Distilling Token-Trained Models into Byte-Level Models

   Zishuo Bao, Jiaqi Leng, Junxiong Wang, Bowen Peng, and Yucheng Lu

   arXiv preprint, 2026

   Abs arXiv

   We propose a low-cost distillation pipeline to convert token-trained LLMs into byte-level models without training from scratch.

   • Method:
     • Progressive Knowledge Distillation to align byte representations with teacher embeddings
     • Byte-Level Supervised Fine-Tuning for end-to-end byte generation
   • Evaluation: Llama, Qwen, and OLMo
   • Result: comparable performance with only  125B bytes of data.

2025

1. NeurIPS 2025

   

   Hardware-aligned Hierarchical Sparse Attention for Efficient Long-term Memory Access

   Xiang Hu, Jiaqi Leng, Jun Zhao, Kewei Tu, and Wei Wu

   In Proceedings of the 39th Conference on Neural Information Processing Systems, 2025

   Abs arXiv Code

   We introduce Hierarchical Sparse Attention (HSA) to give RNNs efficient long-range random access without losing linear complexity.

   • Method: chunk inputs, select top-k chunks, and hierarchically aggregate using token-level relevance; a hardware-aligned kernel keeps it efficient
   • System: Mamba + HSA (RAMba)
   • Result: perfect passkey retrieval up to 64M tokens from 4K pre-training, plus strong downstream gains with near-constant memory.